Searching for information at work: how long do your employees spend looking for answers?

The problem rarely starts with information that does not exist. Usually it is just somewhere else. The technical value is in a manual, the decision in meeting minutes, the current status in the ERP, and the crucial addition arrived by email at some point. Each of these systems makes sense on its own. Together they raise a question nobody answers for the employee: where do I start?

The search itself is unremarkable. Type a search term. Try another one. Click through folders. Open three hits. Each step is harmless, and that is exactly why nobody counts them.

More expensive than the minutes is the uncertainty at the end. Did I search in the wrong place? Is this version the current one? Does the content still apply? Can I work with it? Anyone who cannot answer these questions either carries on with a residual risk or interrupts a colleague. What happens then, we have described in a separate article on knowledge that lives in people. This article is about the road up to that point.

That this road is long is not a guess. In Microsoft’s Work Trend Index 2023, 62 percent of 31,000 respondents said they struggle with too much time spent searching for information. The figure comes from office environments and cannot be transferred to every business. It does not need to be.

Nobody books “14 minutes searching for the current specification” to a cost centre. A single search is too small. Across the whole company, though, it repeats every day. Instead of an international average, a rough calculation of your own is enough:

• 100 employees, each losing 15 minutes per working day to avoidable searching in file shares and systems
• That is 25 working hours per day, or 5,500 hours per year at 220 working days
• At an assumed fully loaded cost of €50 per hour: €275,000 per year

This is not a study but a worked example with its assumptions in the open. Put in your own numbers. The interesting question is then not “Do our people really search that long?” but “How many minutes a day would we have to save for this to matter financially?” The answer is often surprisingly few.

The calculation is deliberately conservative. It includes neither the time of the colleagues who end up being asked anyway, nor the repetition: the same information gets searched for again next week by somebody else, because the first search left no trace anywhere.

The obvious response is: document better, search better. Both help, and both solve only part of the problem. A company can own tens of thousands of files and still have an access problem, because the employee still has to know what the file is called, which terms were used at the time and which version applies. Why more upkeep changes little about that, we have described elsewhere.

The deeper reason: many everyday questions are not search queries. An employee does not ask “Which document contains the material number?” but “Which part do I need for this machine?” Not “Show me all files for the customer Müller” but “What did we promise Müller in the last meeting?” And sometimes: “Why do we actually do this step this way?”

A search engine answers such questions with a list of hits. The answer may be in the third hit, it may only emerge from the third and the seventh together, or it may be in none of them. The work of finding that out stays with the employee. Ten hits in 200 milliseconds are therefore not a success if ten minutes of reading follow.

Perhaps companies are measuring in the wrong place. How many documents are in SharePoint? How many pages does the wiki have? How fast does the search respond? A more interesting question: how long does it take from a technical question to an answer you can act on with confidence?

This time to a reliable answer does not end at the first hit. It ends when four things are clear: this is the information I need. It applies to my case. It is current enough. And I know where it comes from. If any one of these is missing, the search is not over, even if a document is open on the screen.

The metric has an uncomfortable consequence for AI tools. A language model can produce a fluent answer in two seconds, and the time to an answer drops dramatically. The time to a reliable answer only drops if the answer can be checked. Germany’s Federal Office for Information Security (BSI) notes in its publication on generative AI models that fabricated outputs from language models usually appear credible, especially when they come with references. A plausible but wrong answer is therefore worse than a slow correct one, and worse than the honest note that the current status is not known.

The value of an AI knowledge base is therefore not the chat, but three properties. First, it understands the question as it is asked, not just the search term. Second, it returns the passage rather than the hit list, and names the source, so that the four criteria above can be checked in seconds rather than minutes. Third, and this is where good systems part ways with bad ones: if the information is missing, or two sources contradict each other, exactly that has to become visible instead of disappearing into a smooth sentence.

That is how we built hAiner: every answer carries its source, and contradictions and gaps are flagged rather than papered over. Which other types of solution exist, and which suits which kind of knowledge, is set out on our topic page AI knowledge management software.

What AI does not change: knowledge that is written down nowhere cannot be found by any model. If the answer exists only in a colleague’s head, a good knowledge base shows exactly that gap. That is not a shortcoming but the start of the second half of knowledge management: securing experiential knowledge before it leaves the company.

You do not need software for this, just an hour and a stopwatch. A procedure that fits into one afternoon:

1. Collect ten real questions. Not invented ones, but from last week: what was actually searched for or asked? Customer commitments, inspection dimensions, exceptions to a rule, supplier terms.
2. Have two employees search, using the systems you already have. One experienced, one new. Stop the clock not at the first hit, but when the person says: I can work with this.
3. Note the end point. Where was the answer in the end: in a document, in a system, with a colleague, or nowhere? And was the version right?

The result is a small table with ten rows, two times and one end point per question. It says more about the state of your company’s knowledge than the number of documents in your file share. Questions both employees could only resolve through a colleague belong on the list Maike Penz describes in the article on knowledge that lives in people. Questions where the experienced employee was done in two minutes and the new one gave up after twenty show you your access problem.

A single ten-minute search endangers no business. Thousands of small searches a year tie up working time and create something that is hard to express in euros: the feeling of working against your own organization. The aim of knowledge management should therefore not be to store as much as possible, but to keep the road from question to reliable answer short.

How big is your knowledge risk? The knowledge loss check takes five questions and two minutes, no sign-up required.